AMD is #1

Stocks A to Z
1

From https://top500.org/lists/top500/2022/06/

The 59th edition of the TOP500 revealed the Frontier system to be the first true exascale machine with an HPL score of 1.102 Exaflop/s.

The No. 1 spot is now held by the Frontier system at Oak Ridge National Laboratory (ORNL) in the US. Based on the latest HPE Cray EX235a architecture and equipped with AMD EPYC 64C 2GHz processors, the system has 8,730,112 total cores, a power efficiency rating of 52.23 gigaflops/watt, and relies on gigabit ethernet for data transfer.

However, a recent development to the Frontier system has allowed the machine to surpass the 1 exaflop barrier. With an exact HPL score of 1.102 Exaflop/s, Frontier is not only the most powerful supercomputer to ever exist – it’s also the first true exascale machine.

The top position was previously held for two years straight by the Fugaku system at the RIKEN Center for Computational Science (R-CCS) in Kobe, Japan. Sticking with its previous HPL benchmark score of 442 PFlop/s, Fugaku has now dropped to No. 2. Considering the fact that Fugaku’s theoretical peak is above the 1 exaflop barrier, there’s cause to also call this system an exascale machine as well. However, Frontier is the only system able to demonstrate this on the HPL benchmark test.

…

Frontier is the new No. 1 system in the TOP500. This HPE Cray EX system is the first US system with a peak performance exceeding one ExaFlop/s. It is currently being integrated and tested at the ORNL in Tennessee, USA, where it will be operated by the Department of Energy (DOE). It currently has achieved 1.102 Exaflop/s using 8,730,112 cores. The new HPE Cray EX architecture combines 3rd Gen AMD EPYC™ CPUs optimized for HPC and AI with AMD Instinct™ 250X accelerators and Slingshot-11 interconnect.

Although I’m confused about why this is the first “true exascale” machine, since, per https://en.wikipedia.org/wiki/Fugaku_(supercomputer), the Fugaku machine did better 2 years ago:

In June 2020, it achieved 1.42 exaFLOPS (in mixed fp16/fp64 precision) in HPL-AI benchmark making it the first ever supercomputer that achieved 1 exaFLOPS.

3 Likes
2

I’m confused about why this is the first “true exascale” machine, since, per [Wikipedia], the Fugaku machine did better 2 years ago [on the HPL-AI benchmark]

The HPL-AI benchmark is secondary. The primary benchmark, on which the supercomputers are ranked in the Top500 list, is the Linpack benchmark (using 64-bit math). Frontier scores 1.1 exaflops on the Linpack test and 6.9 exaflops on the HPL-AI test (using lower-precision math).

https://cdn.mos.cms.futurecdn.net/7Zb63uiUU4CxPtGmfW2gKW-970…

Notably, this makes Frontier the fastest AI machine also.

“Frontier also now ranks as the fastest AI system on the planet, dishing out 6.88 ExaFlops of mixed-precision performance in the HPL-AI benchmark. […] It appears this system will compete for the AI leadership position with newly-announced AI-focused supercomputers powered by Nvidia’s Arm-based Grace CPU Superchips.”

https://www.tomshardware.com/uk/news/amd-powered-frontier-su…

Also worth noting, Frontier will do leading edge AI (and HPC) with no CUDA in sight. And so will El Capitan, coming next year. The proprietary CUDA moat is steadily being overcome in the supercomputer space, which is good to see.

https://forums.anandtech.com/threads/speculation-sycl-will-r…

1 Like
3

By the way, I love AMD CTO Mark Papermaster’s ambitious and relentless focus on power-efficiency (30x by 2025) — with superstar engineers such as Corporate Fellow Sam Naffziger to make it happen. It has really paid off, as we can see with the big efficiency lead achieved by the latest AMD-based supercomputers:

https://preview.redd.it/l0kq4n6gip291.png?width=960&crop…

PS. In the chart above, note that Selene uses AMD EPYC CPUs combined with Nvidia GPUs.

https://en.wikipedia.org/wiki/Selene_(supercomputer)

2 Likes
4

AMD-Powered Frontier Supercomputer Breaks the Exascale Barrier, Now Fastest in the World
https://www.tomshardware.com/news/amd-powered-frontier-super…

Frontier not only overtakes the previous leader, Japan’s Fugaku, but blows it out of the water — in fact, Frontier is faster than the next seven supercomputers on the list, combined. Notably, while Frontier hit 1.1 ExaFlops during a sustained Linpack FP64 benchmark, the system delivers up to 1.69 ExaFlops in peak performance but has headroom to hit 2 ExaFlops after more tuning.

Frontier also now ranks as the fastest AI system on the planet, dishing out 6.88 ExaFlops of mixed-precision performance in the HPL-AI benchmark.

Additionally, the Frontier Test and Development (Crusher) system also placed first on the Green500, denoting that Frontier’s architecture is now also the most power-efficient supercomputing architecture in the world (the primary Frontier system ranks second on the Top500).

AMD EPYC-powered systems now comprise five of the top ten supercomputers in the world, and ten of the top twenty. In fact, AMD’s EPYC is now in 94 of the Top500 supercomputers in the world,

However, in terms of power efficiency, AMD reigns supreme in the latest Green500 list — the company powers the four most efficient systems in the world, and also has eight of the top ten and 17 of the top 20 spots.

The article also has many more details on the systems specifications, cooling and inner workings if you are interested.

5

Thanks for clearing that up for me guys. AMD really seems to be wiping the floor with the competition in the server space. I wonder how AMD will fare against Tachyum when and if their system comes online later this year and next. From https://www.tachyum.com/solutions/high-performance-computing…

For dedicated, state-of-the-art government-funded Exascale systems, the Prodigy processor scheduled for release in 2022 will directly enable a path to 50-100 AI ExaFLOPS Prodigy-powered machines by 2023, at 3-6x lower cost per ExaFLOPS compared to the announced budget for the latest El Capitan 2 ExaFLOPS system, also due out in 2023.

1 Like
6

For dedicated, state-of-the-art government-funded Exascale systems, the Prodigy processor scheduled for release in 2022 will directly enable a path to 50-100 AI ExaFLOPS Prodigy-powered machines by 2023, at 3-6x lower cost per ExaFLOPS compared to the announced budget for the latest El Capitan 2 ExaFLOPS system, also due out in 2023.

This comparison is worthless. The El Capitan contract includes hardware support and a big chunk of software development and support. That includes porting existing software to the new system. Tachyum needs to clear up what they are including in their comparison. Plus the Top500 list is based on double-precision floating-point. El Capitan (and other AMD supercomputers) zoom way past anything else on mixed-precision AI benchmarks. I’ve probably said this a hundred times on this newsgroup, floating-point code is usually dominated by moving data (and resizing it). AI code involves mixing different precisions. To use the SIMD instructions efficiently, you may have to take single-precision data, pack it into b16 or i8, then unpack the results.

2 Likes